[Iccrg] IETF plenary: fast timescales

[Frank Kelly]

Hi Joel


You asked a great question about time-scales at the IETF Thursday plenary.


I didn't understand the question at the time, but Bob Briscoe did.


Here are my thoughts on the question, as I now understand it.


Imagine a queue where packets are marked if the queue
length is above a threshold and not otherwise,
and imagine that end-systems react to congestion
information much like TCP: they slow down if acknowledgements
show that packets were marked and speed up otherwise.


If the RTT of a flow is long compared with the relaxation time of the queue, 
then the flow's instantaneous sending rate will not be highly correlated with 
the instantaneous queue size, but the average sending rate over periods long 
compared with the RTT will be heavily influenced by the proportion of
packets marked. If this proportion is high, then the flow is slow and the
total number of marks collected for a fixed size file transfer is high.


If the RTT of a flow is short compared with the
relaxation time of the queue, then the flow's
nstantaneous sending rate will be correlated with
the instantaneous queue size: fewer packets will
be sent when the queue size is above the threshold.
In this case the flow has a short enough RTT to be able to take advantage of 
queueing transients.


If I understood your point, it was that queue size fluctuations are becoming 
faster and faster, as a consequence of faster routers. I agree!
See, e.g., http://www.statslab.cam.ac.uk/~frank/smi.html


The above discussion is concerned with signalling congestion,
and how fast end-systems can react, if they want to. How to align incentives 
with congestion signals is of course the tricky issue that Bob is working on - 
http://bobbriscoe.net/projects/refb/


Regards,
Frank Kelly